Human biometric verification and recognition based on feature fusion and computer vision techniques

In this dissertation, several techniques and methods are investigated to solve the human recognition problem. The major contributions in this dissertation include the development of a mathematical framework for discovering human face features, computer vision algorithms to extract features from wavelet transformation and statistical evaluation of features in terms of the feature's ability to discriminate different people. A fuzzy matched filter algorithm is developed to extract face component objects in a front face image. Several neural networks are used to fuse different human biometric features to improve the computer human recognition system performance.; The fuzzy matched filter is not dependent on a constant face template. The very basic relations between the face components are used as the foundation for the fuzzy rules. An optimal decision making method in the sense of maximizing fuzzy membership functions (minimizing uncertainty) is developed in this dissertation to find the face components. A wavelet transformation is introduced as a new method to extract side view face features. Utilizing the wavelet transformation, a side view profile is decomposed as high frequency and low frequency parts. Signal reconstruction, autocorrelation and energy distribution are used to decide a minimum decomposition level in the wavelet transformation without losing side view features. The tie statistic is employed to evaluate the face feature ability to distinguish different people. The good features in terms of discriminating people can be selected by their tie statistic values. Neural networks are used to transfer the feature space to a decision space. The recognition decision can be easily made in decision space when unknown face features are input to the trained neural network.; Several biometric features from human face and hand images are fused to improve computer recognition system performance. This feature fusion is implemented using a neural network. The experimental results show that the algorithms developed in this study achieve satisfactory recognition accuracy.